GPT-6.1 Sol computer use nearly matches GPT-6 Astra at 1/7 the cost on OSWorld 2.0

haider1 · x · 2026-09-30

Evaluator haider reports on GPT-6.1 Sol's computer use: on OSWorld 2.0 it scores 7 points above GPT-6 Sol and finishes just 2 points behind GPT-6 Astra, while costing roughly 1/7 as much per task. He previously noted that on AutomationBench for long-running agent work, GPT-6.1 Sol at medium reasoning beats Opus 5.5 at about 1/3 the cost. Together the results point to near-flagship computer-use performance at a fraction of the price.

Related event: GPT-6.1 Sol Matches Astra at One-Fifth the Cost, API Up to 6x Faster(3 posts)→

Original post →

More from Models

Models channel →